lavfi/dnn: Async Support for TensorFlow Backend

This commit enables async execution in the TensorFlow backend
and adds function to flush extra frames.

The async execution mechanism executes the TFInferRequests on
a separate thread which is joined before the next execution of
same TFRequestItem/while freeing the model.

The following is the comparison of this mechanism with the existing
sync mechanism on TensorFlow C API 2.5 CPU variant.

Async Mode: 4m32.846s
Sync Mode: 5m17.582s

The above was performed on super resolution filter using SRCNN model.

Signed-off-by: Shubhanshu Saxena <shubhanshu.e01@gmail.com>

libavfilter/dnn/dnn_backend_tf.c

@@ -38,7 +38,6 @@
 #include "dnn_io_proc.h"
 #include "dnn_backend_common.h"
 #include "safe_queue.h"
-#include "queue.h"
 #include <tensorflow/c/c_api.h>
 
 typedef struct TFOptions{
@@ -59,6 +58,7 @@ typedef struct TFModel{
     TF_Status *status;
     SafeQueue *request_queue;
     Queue *inference_queue;
+    Queue *task_queue;
 } TFModel;
 
 /**
@@ -75,7 +75,7 @@ typedef struct TFInferRequest {
 typedef struct TFRequestItem {
     TFInferRequest *infer_request;
     InferenceItem *inference;
-    // further properties will be added later for async
+    DNNAsyncExecModule exec_module;
 } TFRequestItem;
 
 #define OFFSET(x) offsetof(TFContext, x)
@@ -89,6 +89,7 @@ static const AVOption dnn_tensorflow_options[] = {
 AVFILTER_DEFINE_CLASS(dnn_tensorflow);
 
 static DNNReturnType execute_model_tf(TFRequestItem *request, Queue *inference_queue);
+static void infer_completion_callback(void *args);
 
 static void free_buffer(void *data, size_t length)
 {
@@ -886,6 +887,9 @@ DNNModel *ff_dnn_load_model_tf(const char *model_filename, DNNFunctionType func_
             av_freep(&item);
             goto err;
         }
+        item->exec_module.start_inference = &tf_start_inference;
+        item->exec_module.callback = &infer_completion_callback;
+        item->exec_module.args = item;
 
         if (ff_safe_queue_push_back(tf_model->request_queue, item) < 0) {
             av_freep(&item->infer_request);
@@ -899,6 +903,11 @@ DNNModel *ff_dnn_load_model_tf(const char *model_filename, DNNFunctionType func_
         goto err;
     }
 
+    tf_model->task_queue = ff_queue_create();
+    if (!tf_model->task_queue) {
+        goto err;
+    }
+
     model->model = tf_model;
     model->get_input = &get_input_tf;
     model->get_output = &get_output_tf;
@@ -1061,7 +1070,6 @@ static DNNReturnType execute_model_tf(TFRequestItem *request, Queue *inference_q
 {
     TFModel *tf_model;
     TFContext *ctx;
-    TFInferRequest *infer_request;
     InferenceItem *inference;
     TaskItem *task;
 
@@ -1074,23 +1082,14 @@ static DNNReturnType execute_model_tf(TFRequestItem *request, Queue *inference_q
     tf_model = task->model;
     ctx = &tf_model->ctx;
 
-    if (task->async) {
-        avpriv_report_missing_feature(ctx, "Async execution not supported");
+    if (fill_model_input_tf(tf_model, request) != DNN_SUCCESS) {
         return DNN_ERROR;
-    } else {
-        if (fill_model_input_tf(tf_model, request) != DNN_SUCCESS) {
-            return DNN_ERROR;
-        }
+    }
 
-        infer_request = request->infer_request;
-        TF_SessionRun(tf_model->session, NULL,
-                      infer_request->tf_input, &infer_request->input_tensor, 1,
-                      infer_request->tf_outputs, infer_request->output_tensors,
-                      task->nb_output, NULL, 0, NULL,
-                      tf_model->status);
-        if (TF_GetCode(tf_model->status) != TF_OK) {
-            tf_free_request(infer_request);
-            av_log(ctx, AV_LOG_ERROR, "Failed to run session when executing model\n");
+    if (task->async) {
+        return ff_dnn_start_inference_async(ctx, &request->exec_module);
+    } else {
+        if (tf_start_inference(request) != DNN_SUCCESS) {
             return DNN_ERROR;
         }
         infer_completion_callback(request);
@@ -1127,6 +1126,83 @@ DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNExecBaseParams *
     return execute_model_tf(request, tf_model->inference_queue);
 }
 
+DNNReturnType ff_dnn_execute_model_async_tf(const DNNModel *model, DNNExecBaseParams *exec_params) {
+    TFModel *tf_model = model->model;
+    TFContext *ctx = &tf_model->ctx;
+    TaskItem *task;
+    TFRequestItem *request;
+
+    if (ff_check_exec_params(ctx, DNN_TF, model->func_type, exec_params) != 0) {
+        return DNN_ERROR;
+    }
+
+    task = av_malloc(sizeof(*task));
+    if (!task) {
+        av_log(ctx, AV_LOG_ERROR, "unable to alloc memory for task item.\n");
+        return DNN_ERROR;
+    }
+
+    if (ff_dnn_fill_task(task, exec_params, tf_model, 1, 1) != DNN_SUCCESS) {
+        av_freep(&task);
+        return DNN_ERROR;
+    }
+
+    if (ff_queue_push_back(tf_model->task_queue, task) < 0) {
+        av_freep(&task);
+        av_log(ctx, AV_LOG_ERROR, "unable to push back task_queue.\n");
+        return DNN_ERROR;
+    }
+
+    if (extract_inference_from_task(task, tf_model->inference_queue) != DNN_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "unable to extract inference from task.\n");
+        return DNN_ERROR;
+    }
+
+    request = ff_safe_queue_pop_front(tf_model->request_queue);
+    if (!request) {
+        av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
+        return DNN_ERROR;
+    }
+    return execute_model_tf(request, tf_model->inference_queue);
+}
+
+DNNAsyncStatusType ff_dnn_get_async_result_tf(const DNNModel *model, AVFrame **in, AVFrame **out)
+{
+    TFModel *tf_model = model->model;
+    return ff_dnn_get_async_result_common(tf_model->task_queue, in, out);
+}
+
+DNNReturnType ff_dnn_flush_tf(const DNNModel *model)
+{
+    TFModel *tf_model = model->model;
+    TFContext *ctx = &tf_model->ctx;
+    TFRequestItem *request;
+    DNNReturnType ret;
+
+    if (ff_queue_size(tf_model->inference_queue) == 0) {
+        // no pending task need to flush
+        return DNN_SUCCESS;
+    }
+
+    request = ff_safe_queue_pop_front(tf_model->request_queue);
+    if (!request) {
+        av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
+        return DNN_ERROR;
+    }
+
+    ret = fill_model_input_tf(tf_model, request);
+    if (ret != DNN_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to fill model input.\n");
+        if (ff_safe_queue_push_back(tf_model->request_queue, request) < 0) {
+            av_freep(&request->infer_request);
+            av_freep(&request);
+        }
+        return ret;
+    }
+
+    return ff_dnn_start_inference_async(ctx, &request->exec_module);
+}
+
 void ff_dnn_free_model_tf(DNNModel **model)
 {
     TFModel *tf_model;
@@ -1135,6 +1211,7 @@ void ff_dnn_free_model_tf(DNNModel **model)
         tf_model = (*model)->model;
         while (ff_safe_queue_size(tf_model->request_queue) != 0) {
             TFRequestItem *item = ff_safe_queue_pop_front(tf_model->request_queue);
+            ff_dnn_async_module_cleanup(&item->exec_module);
             tf_free_request(item->infer_request);
             av_freep(&item->infer_request);
             av_freep(&item);
@@ -1147,6 +1224,14 @@ void ff_dnn_free_model_tf(DNNModel **model)
         }
         ff_queue_destroy(tf_model->inference_queue);
 
+        while (ff_queue_size(tf_model->task_queue) != 0) {
+            TaskItem *item = ff_queue_pop_front(tf_model->task_queue);
+            av_frame_free(&item->in_frame);
+            av_frame_free(&item->out_frame);
+            av_freep(&item);
+        }
+        ff_queue_destroy(tf_model->task_queue);
+
         if (tf_model->graph){
             TF_DeleteGraph(tf_model->graph);
         }

libavfilter/dnn/dnn_backend_tf.h

@@ -32,6 +32,9 @@
 DNNModel *ff_dnn_load_model_tf(const char *model_filename, DNNFunctionType func_type, const char *options, AVFilterContext *filter_ctx);
 
 DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNExecBaseParams *exec_params);
+DNNReturnType ff_dnn_execute_model_async_tf(const DNNModel *model, DNNExecBaseParams *exec_params);
+DNNAsyncStatusType ff_dnn_get_async_result_tf(const DNNModel *model, AVFrame **in, AVFrame **out);
+DNNReturnType ff_dnn_flush_tf(const DNNModel *model);
 
 void ff_dnn_free_model_tf(DNNModel **model);
 

libavfilter/dnn/dnn_interface.c

@@ -48,6 +48,9 @@ DNNModule *ff_get_dnn_module(DNNBackendType backend_type)
     #if (CONFIG_LIBTENSORFLOW == 1)
         dnn_module->load_model = &ff_dnn_load_model_tf;
         dnn_module->execute_model = &ff_dnn_execute_model_tf;
+        dnn_module->execute_model_async = &ff_dnn_execute_model_async_tf;
+        dnn_module->get_async_result = &ff_dnn_get_async_result_tf;
+        dnn_module->flush = &ff_dnn_flush_tf;
         dnn_module->free_model = &ff_dnn_free_model_tf;
     #else
         av_freep(&dnn_module);